Modernizing travel applications in the age of gen AI

The technology will have a massive impact on the sector—so keeping pace is vital.

From recommending destinations to optimizing ticket pricing to forecasting equipment maintenance to creating more efficient and eco-friendly routes, the travel sector is on the cusp of a transformation driven by generative AI.

In spite of questions around costs and implementation risks, travel and transportation companies are increasingly adopting gen AI to revolutionize business processes and enhance customer experience (CX). For instance, Expedia leverages AI multi-agent solutions to streamline booking processes and provide superior customer support. And Air India invested $200 million in OpenAI’s ChatGPT to automate customer service functions, helping agents respond to complex situations more effectively.

Businesses in this sector that fail to keep pace with gen AI initiatives risk losing out to more progressive competitors. To that end, we offer a practical guide with actionable use cases demonstrating ways to augment existing travel systems using AI, emphasizing data enhancement, risk reduction, and cost management for successful adoption.

Simplify the UI to improve CX

Gen AI is set to simplify user interfaces significantly. Key differentiators will be how information is delivered, how tasks are performed, and the extent of personalization. Ideally, there should be just one persona—either a multi-modal chat assistant or voice assistant, or a virtual reality assistant.

To bring this about, organizations must consolidate their services and use large language model (LLM) wrappers to modernize their applications. This reduces the number of interface windows and helps store user activities in one place for better personalization.

To grasp the potential for improved CX here, consider how many applications an airline customer might interact with:

• Planning and booking
• Pre-flight preparation (online check-in, travel document submission, etc.)
• Onboarding
• In-flight experience (Wi-Fi access, in-flight purchases, etc.)
• Off-boarding (disembarkation, baggage collection, customs clearance, etc.)
• Post-flight activities (rewards, feedback/reviews)

Adopt new design thinking

The advantages of a single, consolidated UI for all these activities are immediately evident. Multi-agent AI systems are transforming enterprise operations by enabling AI agents to interact not only with humans but also independently among themselves. Such interactions require a more integrated approach to business processes, breaking down silos between different systems, and a long-term plan:

• Enhanced efficiency and usability. A UI overhaul is needed under which users can enter free-flow text instead of navigating through menus. Responses will be generated based on queries or requests. The UI should support seamless question-answering across diverse data types (text, tables, images, etc.) and enable multi-modal embeddings for images, videos, voice, charts, and maps. These UI design considerations should be reflected in system messages when implementing tools. For example, if a response requires a table, this instruction should be clearly defined during the LLM tool setup. Special care should be taken to ensure a smooth UI transition.
  
  
• Interconnected systems. Agentic workflow is poised to transform software as a service, or SaaS, into AaaS (Agent-as-a-Service), featuring two layers: CRUD services and AI-driven orchestration. To modernize existing APIs, applications, or systems, businesses can create agents by wrapping them as LLM tools with clearly defined functionalities, arguments, and return types. These tools can be grouped under an agent. For example, flight search, booking, update, and cancel can be different tools grouped under a flight agent. A common pattern is to use a primary (or “supervisor”) agent to coordinate and delegate tasks between multiple agents. It’s worth mentioning that while orchestration will be managed mostly by LLMs, the services they consume to execute tasks will be built by integrating systems or applications using middleware.
  
  For example, if a user requests, “Book a hotel for my next trip,” the application would first fetch user information and pass it to the primary assistant. The primary assistant would then fetch the trip details using the booking agent, which queries a database or uses an API to get the data. The primary agent would extract the destination and the check-in and check-out dates and would pass this information to the hotel agent to search for hotels. Available hotel results would be displayed to the end user for booking confirmation. The primary assistant can personalize the information or offer based on user activities, preferences, or other digital footprints.
  
  Thus, all orchestration is handled by LLM-based agents, unlike current rule-based orchestration workflows. Performing such complex tasks through a single command is almost impossible with rule-based orchestration. The flip side is that due to unlimited possibilities, the behavior can be inconsistent and potentially misused. This is why having granular agents helps clarify the intent to ensure consistent behavior, such as updating a flight, which means first canceling the existing one.
  
  
• Transformative potential. When modernizing existing applications, organizations should seize every opportunity to leverage AI and LLMs to create a network of intelligent tools that can revolutionize business operations. These tools can enhance areas such as personalization, security, multi-modality, sustainability, inclusivity, and self-healing capabilities. This process may involve fine-tuning models, which requires appropriate datasets and budget considerations. Therefore, a cost-benefit assessment is essential before embarking on any AI-based transformation journey.

Build your own ecosystem to avoid model collapse

Modernization goes beyond merely wrapping existing applications or services with LLMs or introducing generative AI for insights and content generation. It also involves the integration of business processes, which means integrating applications and data sources. These data sources may be internal or external, and external sources may be in the public domain or within vendor or partner networks. Reaching a consensus on integrating these diverse sources is challenging due to legal, ethical and political barriers. This requires meticulous reasoning, business case creation, and ROI definition, all underpinned by a unified vision.

Moreover, resolving data integration issues is just the beginning; true efficiency and performance depend on data quality. For example, personalization involves tailoring promotional content, recommendations, and services to individual customers. The accuracy of personalization depends on how data from different domains is collected, correlated and consumed, along with the selection of an appropriate AI model.

Customers’ digital footprint in the public domain is increasingly polluted by the rapid growth of AI-generated content, a phenomenon known as “model collapse.” To mitigate this issue, organizations must make a committed and continuous effort to capture the digital footprint of customers at every stage of their journey (as shown in the following figure), in collaboration with data providers and partners. This proactive approach ensures the integrity and accuracy of the data used for personalization and other AI-driven initiatives.

Figure 1

A cohesive strategy is a must

Modernization is a long-term, organization-wide initiative that requires strong leadership buy-in and sponsorship. Without a clear vision, defined ROI, and thorough planning, navigating through modernization or transformation efforts can be challenging. Cost is often the biggest barrier, especially since implementing agentic workflows is generally more expensive than rule-based workflows. Therefore, it’s crucial to understand when and why to use agentic workflows.

The primary advantage of agentic workflows is their ability to execute tasks in a smarter, more autonomous and efficient manner using natural language. The implementation will be most impactful when it directly benefits a large number of customers or employees and covers wide, interconnected functionalities.

It’s also important to determine when to use local LLMs; cloud-based models; or fine-tuned existing models. Typically, creating a custom model from scratch is unnecessary unless the organization is targeting a niche and unique area as a key differentiator. For reasoning and action tasks, simple local or cloud-based LLMs can be used. Local LLMs are generally cheaper and have lower latency compared to cloud-based LLMs, but they come with trade-offs in scalability and maintenance.

In some cases, such as creating personalized offerings in the travel industry, fine-tuning an existing gen AI model works. Fine-tuning enhances the model’s ability to understand and predict customer preferences, enabling more precise personalization. However, the appropriateness of fine-tuning depends on the availability of data and a thorough cost-benefit analysis.

Mitigating risk: crawl, walk, run

As the gen AI landscape continues to evolve, unforeseen issues are sure to arise. The fact is, this new field is full of risk. To mitigate this risk, a meticulous and gradual rollout plan is essential, along with thorough testing. Initially, the new modernized interface should be rolled out to internal customers with proper prioritization of functionalities. For example, an LLM-enabled, natural language-driven planning and booking system can first be introduced to customer service agents to handle customer requests. Once the solution proves satisfactory, it can be extended to partners such as external booking agents. Finally, a stable, tested, consistent, and secure application can be made available to end users.

A holistic framework is essential to mitigate risks associated with modernization or transformation initiatives. This framework should address four root causes:

• Poor data quality
• Inadequate risk control
• Unclear business value
• Escalating costs

While implementing this framework, it’s imperative to balance strategic and tactical needs, efficiency and growth objectives, and current and future capabilities.

Travel and the agentic future

As new chips and AI models emerge, the agentic workflow framework will evolve alongside them. However advanced these models may be, they can still fall short in addressing emerging challenges due to insufficient contextual training data. This limitation highlights the critical role of human expertise.

Skilled professionals are vital for interpreting AI outputs, making informed decisions, and implementing robust application upgrades. Thus, cultivating an AI-literate workforce through comprehensive training and change management is essential. This approach ensures that employees are not only adept at using AI tools but also adaptable to the ever-evolving technological landscape.